Source-grounding is the practice of forcing an AI answer to come from specific passages of source text — and attaching a citation to each claim — instead of letting the model answer from its own parametric memory. A grounded answer can always point to the sentence it came from; an ungrounded one cannot.
Why it matters
A language model trained on the open web carries a vast, blurry impression of facts in its weights. Ask it a question and it will happily answer from that memory, but it has no way to tell you which document, page, or sentence the answer came from — because there often isn't one. That is exactly how hallucinations creep in: plausible-sounding text with nothing real behind it.
Source-grounding flips the contract. The answer is no longer "what the model remembers about this topic" but "what these specific passages actually say." That makes answers auditable: you can check the citation, see the original wording, and decide for yourself whether the model read it correctly.
How it works
Grounding usually has two halves:
- Retrieval. Retrieval-augmented generation pulls the most relevant passages from the document and places them in the model's context, so it has real text to work from.
- Constraint and attribution. The model is instructed to answer only from those passages, and each claim is tied back to the passage that supports it with a citation.
The crucial test is what happens when the passages do not support a claim. A truly grounded system drops or flags that claim rather than guessing.
Sidenote treats grounding as the default, not a feature: every answer is built from passages retrieved out of the document you are reading, each citation scrolls to the exact source, and any claim that cannot be matched to a real passage is dropped before it ever reaches you.